On this page
On 24 September 2026 (AEST), the Prime Minister, the Hon Anthony Albanese MP, announced details of an incident in which an AI agent from OpenAI infiltrated an Australian Government website – the Medicare Statistics Reporting Service administered by Services Australia.i Further investigation has confirmed that no sensitive personal data was accessed, and the issue has been remediated by taking down the Medicare Statistics Reporting Service and moving the data to a more secure environment.
Back to topWhat happened
On 18 June 2026, an OpenAI agent1 undertook unauthorised activity on Services Australia’s Medicare Statistics Reporting Service. The Medicare Statistics Reporting Service was an unauthenticated, standalone and publicly accessible service, not integrated with other Services Australia systems. The AI model was an internal OpenAI model that did not include the full set of safeguards 2 normally associated with a publicly released model.ii While undergoing evaluation, the agent was set a task to research government spending per person on medicines for skin conditions in Victorian communities.iii
The Australian Signals Directorate’s (ASD’s) Australian Cyber Security Centre (ACSC) conducted an investigation into the June 2026 exploitation of Services Australia’s Medicare Statistics Reporting Service portal. The investigation focused on what data was accessed, if additional systems were affected, and if human input caused the activity. The investigation revealed:
- That the OpenAI internal AI model agent was not appropriately contained, lacked appropriate and timely human oversight or intervention.
- The agent initially attempted to access non-public medical statistical data from the Australian Institute of Health and Welfare (AIHW); however, attempts were unsuccessful. Subsequently, the agent targeted the Services Australia Medicare Statistics Reporting Service portal and successfully accessed public and non-public data without authorisation.
- No evidence of the agent abusing credentials or vulnerabilities to move laterally across the Services Australia network to access more sensitive data on the network, such as Personally Identifiable Information.
- No indication that unauthorised activity was caused by human input.
- No evidence of lateral movement, nor sophisticated anti-forensic activity by the agent.
The OpenAI incident affecting Services Australia represents a realised AI safety risk and was the first documented example of misaligned agents affecting Australian Government infrastructure.
OpenAI became aware of this activity on 11 August 2026, following human review of AI model activity by its security team.iv This OpenAI review was in response to a previous incident in July 2026 when a different OpenAI model broke out of testing and breached the cyber defences of a United States (US) AI company, Hugging Face (the OpenAI/Hugging Face incident).v
OpenAI informed Services Australia on 10 September 2026 (AEST) (9 September, UTC) via email that:
- a security vulnerability had been identified
- an OpenAI model had identified a way to carry out instructions on the server through the public reporting interface, without a private account or password
- the model had been able to access this to read portions of the internal program files and settings, obtain a list of files, and create and read back a small test file on the server, and
- OpenAI had not found any evidence that the model had accessed patient-level records, personal information or personal credentials; deleted data; or established ongoing access.3
The information from OpenAI at this initial stage did not cover full details, including that the activity involved OpenAI performing model testing, or was an activity by an AI agent with no human control. It also contained an incorrect URL which complicated validation.
OpenAI notified Services Australia through a public-facing email (publicised on the Services Australia website), established for reporting potential cyber security system risks affecting Medicare, Centrelink, Child Support and myGov online systems.vi This reporting mechanism is part of Services Australia’s Vulnerability Disclosure Program4 , which supports compliance with the PSPF and is part of the agency’s broader approach to identify, assess and manage cyber vulnerabilities.vii The website publicising the mailbox notes that the mailbox is not intended to report cyber attacks and redirects those reports to the ‘Report a cybercrime’ form on cyber.gov.au.
The vulnerability identified by the agent, and the resulting unauthorised access, were not known to the Australian Government prior to OpenAI’s disclosure to Services Australia on 10 September 2026.
Following disclosure of the OpenAI incident affecting Services Australia, OpenAI advised that its model had also attempted to access to the Australian Institute of Health and Welfare website without success, and had conducted activity against systems operated by Victoria’s Department of Health, New South Wales Bureau of Crime Statistics and Research, and the New South Wales National Parks and Wildlife Service web app. Disclosures from AI companies and independent research has suggested these instances may reflect a broader series of rogue AI-agent activity globally.viii
In response to these incidents, OpenAI has committed in public and directly to the Australian Government that it will strengthen research safeguards, including via additional network restrictions and expanded monitoring. Specific to Australia, OpenAI has publicly affirmed it will provide dedicated support for affected agencies, funding and support to strengthen cyber defences, and establish an Australian taskforce to develop practical policy recommendations for managing emerging risks from increasingly capable AI agents.ix
Back to topWhat the Australian Government did
On 11 September 2026, the email notification from OpenAI was reviewed by a Services Australia agency officer – a day lag resulting from non-continuous monitoring of the public mailbox. In accordance with Services Australia’s cyber security vulnerability management policy, the OpenAI disclosure was referred internally for technical investigation to verify the vulnerability and assess whether there were any impacts to other systems.
On 14 September 2026, the OpenAI disclosure, technical verification and preliminary findings were escalated to Services Australia’s Chief Information Security Officer and subsequently referred to Services Australia’s 24/7 Cyber Security Operations Centre which continued investigative analysis overnight.
On 15 September 2026, following technical investigation and assessment of the vulnerability’s severity, the risks associated with continued public access, and available containment options, Services Australia took the Medicare Statistics Reporting Service offline, declared a cyber security incident, and notified ASD’s ACSC.5 At this point in time, the available information from OpenAI still did not disclose that this event involved a model testing activity undertaken by OpenAI itself or that it involved a non-directed activity by an AI agent.
On 17 September 2026, Services Australia briefed the Office of the Minister for Government Services, informing of the confirmed vulnerability, the decision to take the Medicare Statistics Reporting Service offline, ongoing investigations, and engagement with ASD and OpenAI. Over the course of subsequent days, ASD undertook further investigations and provided briefing to relevant Ministers and ministerial offices, through which the National Cyber Security Coordinator (the Coordinator) was advised of the incident on 19 September 2026.
On 20 September 2026, OpenAI informed ASD verbally that the incident involved its own model testing, and it was an AI agent that had accessed information not available to the public. Following this disclosure, ASD notified the Minister for Government Services, Minister for Cyber Security and the National Cyber Security Coordinator. The National Cyber Security Coordinator commenced discussions with OpenAI the same day.
On 22 September 2026, OpenAI participated in its first technical exchange with ASD, Services Australia, the AISI, and the NOCS. The technical information that ASD needed to support the forensic investigation was not provided by OpenAI until 24 September 2026.
On 24 September 2026, the Prime Minister announced the incident and the establishment of the Review. The Review’s terms of reference were published on the Department of the Prime Minister and Cabinet’s (PM&C) website on 24 September 2026.x
Footnotes
- An AI agent refers to a system or program that can autonomously perform tasks on behalf of a user or another system.Return to footnote 1 ↩
- Model evaluation without safeguards is a normal step for developers to establish a baseline of a model’s true capabilities, risks and biases before safety or alignment techniques are applied.Return to footnote 2 ↩
- Patient level records and personal information are subject to stronger cyber security protections and would not have been accessible through the same means.Return to footnote 3 ↩
- As mandated under PSPF Requirement 0115.Return to footnote 4 ↩
- ASD classified the incident as Category 4 – a ‘low-level malicious attack’, affecting federal government. The vulnerability that was exploited was assessed as unsophisticated and reports indicated no sensitive material was accessed (such as personal information).Return to footnote 5 ↩